Automobile Insurance Claim Fraud Detection using Machine Learning Techniques

Abstract

"As of today, Automobile insurance companies currently use a manual process to investigate fraudulent claims. Recent studies have shown that manual process has failed, and the companies are losing billions of profit every year. The manual investigation involves a lot of people with expertise and money too. The investigation can be time-consuming for the insurance company and the insured and can affect the loyal customer base when a legitimate claim has been investigated as fraud. Even if the investigation is done thoroughly, the accuracy of the process is very low. The research goal is to design and develop an Automobile Claim Fraud Detection System to address the problem to improve the accuracy and simplify fraud investigations for both customers and insurers. The proposed solution has been done using the best available machine learning techniques reviewed in the literature. An ensemble learning technique has been utilized to combine the results of 4 selected classification algorithms to achieve a better performance than each induvial algorithm. As discussed in the literature review, XG Boost, Cat Boost, Ada Boost, and Logistic Regression algorithms were chosen as the most performing classifiers with a voting classifier, a well-known ensemble technique used. Random Oversampling has been used to solve the skewed data problem, and the Chi-squared test has been used to identify the best subset of features to get a meaningful claim fraud prediction model, which has been an untouched part in recent years. The proposed system was tested using several metrics to cover all areas of the study. Domain experts assessed the research's scope and importance, while technical experts assessed the system's design and implementation. According to the evaluation results, the proposed system outperformed the existing systems significantly in accuracy and recall metrics."